Mail   |   LinkedIn   |   Google Scholar   |   GitHub   |   CV

Last updated: March 2026

I recently completed my MSc in AI at the University of Amsterdam, where I enjoyed doing research and contributed to several research papers. My thesis developed a local, affordance-centric world model for zero-shot model predictive control in articulated object manipulation, supervised by Dr. Andrii Zadaianchuk and Prof. Efstratios Gavves.

I enjoy diving deep into recent advancements in AI while always continuing to solidify my understanding of the fundamentals. Currently, I am also working as an AI Software Developer at LeerLevels, integrating generative AI into the platform. Outside of work, I enjoy hiking, playing squash, and nerding out over Formula 1.

I am currently seeking new opportunities as an AI Engineer or AI Research Engineer, so feel free to reach out!

We introduce Morpheus, the first benchmark for evaluating video generation models on physical reasoning using real physical experiments. Our findings reveal that even with advanced prompting and video conditioning, current models struggle to encode physical principles despite generating aesthetically pleasing videos.

In this paper, we introduce HIVE, an interactive dashboard for exploring and interpreting hyperbolic embeddings. It supports several interaction modes and allows switching between two projection methods: CO-SNE and HoroPCA.

In this paper, we reproduce the findings that show that SotA neural IR models struggle with negation, often performing at/or below random ranking baselines. We extend their work by evaluating new SotA models and assessing how well their understanding of negation generalizes across datasets.

In this study, we assessed the reproducibility of the paper “Learning Fair Graph Representations Via Automated Data Augmentations”. We were able to partially reproduce one of the claims made by the original authors and fully reproduce the other two. Beyond verifying the original work, we extended the framework to assess its usability for another downstream task, specifically link prediction. We found that it significantly outperforms baselines on one fairness metric, while performing comparably on the other.

We present a novel use of conformal prediction for time series forecasting that incorporates time series decomposition, allowing us to customize employed methods to account for the different exchangeability regimes underlying each time series component.

Source: adapted from Alexander Timans' fork of Dharmesh Tailor's fork of Leonid Keselman's fork of John Barron's website.